P08-02

Machine learning models for predicting cross-reactivity of beta-lactam antibiotic allergy

Shoki HOSHIKAWA *¹, Keisuke FUKUI², Yukiko KARUO¹, Atsushi TARUI¹, Masaaki OMOTE¹, Kazuyuki SATO¹, Kentaro KAWAI¹

¹Faculty of Pharmaceutical Sciences, Setsunan University
²Nozaki Tokushukai Hospital
( * E-mail: shoki.hoshikawa@setsunan.ac.jp )

In clinical practice, not a few patients have a penicillin allergy, and pharmacists use cross-allergy tables for beta-lactam antibiotics to propose changes to the medication. However, even when switching to a drug that is generally contraindicated, there are cases where no allergic symptoms occur, and there are things that cannot be explained using the cross-allergy table for beta-lactam antibiotics. In clinical practice, there are cases where, from the perspective of treating infectious diseases, it is necessary to reluctantly change to a drug from a different class with a broad spectrum of action that is not a beta-lactam antibiotic, but there is a need in the field to refrain from using drugs with as broad a spectrum of action as possible from the perspective of drug resistance.
Therefore, we investigated the possibility of using AI to suggest drugs that can be used in the same class of antimicrobial agents as a substitute for narrow-spectrum beta-lactam drugs. Specifically, we evaluated the prediction ability of a leave-one-out approach to evaluating the combination of two drugs (in four categories: recommended, caution, principal contraindication, and contraindication) in 35 beta-lactam antimicrobial agents. Here, three fingerprints (MACCS, Morgan, and Topological fingerprints) and four machine learning methods (SVM, lightGBM, Random Forest, and k-nearest neighbor) were used for the study. As a result, the highest accuracy of 0.98 was achieved when predicting the recommended drug combinations using Morgan fingerprints with lightGBM. On the other hand, the lowest accuracy of 0.89 was obtained when predicting the drug combinations of the caution category using MACCS fingerprints with SVM. We then focused on the compounds for which the prediction did not work well and evaluated the reasons from the perspective of structural similarity. We also evaluated the relationship between the combinations in each category and structural similarity using Tanimoto similarity of fingerprints. Drug combinations that are recommended by AI while having low similarity may be useful in clinical drug selection.